Ann An empirical review was published A paper published in the journal Patterns claims that artificial intelligence systems are developing skills in deception, and argues that the ability to lie and manipulate is something AI is already using to trick humans for its own gain.
AI deception is not a bug or glitch in a particular system. AI deception capabilities have been reported in special-purpose systems designed to serve humans, as well as general-purpose large-scale language models. This raises questions about the future of AI development and whether we can trust such systems.
As a result, the study’s summary section presented possible solutions to detect AI deception, noting that “regulatory frameworks should apply strict risk assessment requirements to AI systems capable of deception” and that “policymakers should implement bot or no bot laws.” The study further suggested that “policymakers should prioritize funding for related research, including tools to detect AI deception and reduce the deceptive nature of AI systems.”
AI Development and Operation
CNN’s Jake Tapper A quick interview with Geoffrey HintonThe computer scientist and cognitive psychologist is scheduled to give a talk on this very issue in 2023.
Tapper asked Hinton whether AI would have the ability to “figure out how to manipulate or even kill humans,” to which Hinton replied, “If an AI becomes much smarter than humans, it will be very good at manipulation because it will have learned it from humans. And there are very few examples of more intelligent things being controlled by less intelligent things.”
Hinton stressed that a major problem in the future will be AI’s ability to manipulate humans, and he expressed pessimism about whether the rapid development of AI can be stopped.
BlazeNews reached out to Samuel Hammond, senior economist at the Foundation for American Innovation, who said, “Large-scale language models like ChatGPT are trained on virtually all the text in the world, including billions of examples of humans lying to each other, so it’s not surprising they’re capable of lying as well.”
Hammond is Well-known cases In an early test of GPT-4, the team “asked a TaskRabbit worker to complete a Captcha,” which is a program or system specifically designed to distinguish between human and machine input. But when the worker jokingly asked GPT-4 if it was a robot, the AI ”inferred that he should lie and claim to be visually impaired,” and thus avoided the Captcha.
“An example of an emerging capability is whether an AI can lie. Lying requires a mental model of what other people are thinking — what psychologists call ‘theory of mind.’ As LLM scales, we’ll see AI improve significantly in tasks that require theory of mind, sometimes even surpassing human performance.”
“LLMs are not trained to intentionally lie; they are simply trained to complete reams of text and follow user instructions. The ability to lie emerges naturally as general intelligence and reasoning abilities increase. As larger models are trained, AI researchers need to test deceptiveness and related abilities because they do not know in advance what new abilities will emerge,” Hammond said.
While building the technological infrastructure to “test deceptiveness and related abilities” is a lofty goal, it is still not clear how AI could develop a “theory of mind” or metacognition.
“The study’s lead author, Dr. Peter S. Park, a Postdoctoral Research Fellow in AI Existential Safety at MIT, Said“AI developers do not understand with any certainty the causes of undesirable AI behavior, such as deception.”
“But generally speaking, we think that AI deception occurs because a deception-based strategy turns out to be the best way to get a particular AI to perform well on a training task; deception helps the AI achieve its goals,” Park added.
The study carefully evaluated several AI systems and found that many of them have developed deceptive capabilities through the training process. The systems analyzed ranged from AI game playing to more general models used in economic and safety testing environments.
One A striking example of deception The study cited concerned Meta’s CICERO, an AI developed specifically to play the game Diplomacy. Although the AI was trained to behave honestly and maintain alliances with human players, CICERO often deployed deceptive strategies to win the game.
CICERO created false alliances when it gave it an advantage in gameplay performance, leading researchers to conclude that Meta’s AI programs were “masters of deception.”
“Despite Meta’s efforts, CICERO proved to be a master liar,” the researchers wrote. “Not only did it betray other players, but it also engaged in deliberate deception, planning ahead to form false alliances with human players to trick them and leave them vulnerable to attack.”
The philosophical dilemma of AI deception
Former Google CEO Eric Schmidt Recently mentioned “Truth” is a difficult problem to solve, especially in the age of AI, and the “fundamental problem” with truth, he noted, is that we cannot work out whether people really believe what they write or say, or whether they are simply being paid or incentivized to say that they believe what they write or say.
“I personally think this is unsolvable. That’s a depressing answer. … The reason it’s unsolvable is because you have to rely on critical thinking people. And there’s a lot of evidence…” [that] “People are very gullible, especially to videos that are charismatic or charismatic,” Schmidt said.
Schmidt seems to suggest that AI-generated lies rely on the human population lacking the ability for critical thinking: Without evidence or justification for what humans believe to be true, we are susceptible to being fooled by intelligent computer systems that claim what they are saying is true, even if it is in fact a lie.
James PoulosThe host of the “Zero Hour with James Poulos” podcast told Blaze News: “This relates to Pontius Pilate’s question, ‘What is truth?’ In other words, what reality must be like so that truth alone is sufficient grounds for a truth-teller or thing to be deemed trustworthy.”
“The Christian answer to this question is Christ’s statement: ‘I am the truth’ – in other words, truth is not ‘what’ but ‘who,'” Poulos added. “Those who reject or seek to evade Christ’s answer to the question of what makes truth trustworthy must come up with a different answer – a difficult and demanding task indeed, and one that philosophers have struggled to convince themselves and each other of for millennia.”
Poulos appears to be referring to epistemology, or the theory of knowledge. Many philosophical schools claim there is a justification for how to arrive at fully justified beliefs, but there is no consensus on the issue. And as AI continues to develop, it may become more difficult to establish a basis for truth.
The impact of AI on truth is a growing concern for many. Christian Scholar’s Review report Eric Horwitz, Microsoft’s chief science officer, said: The AI writes: We are gradually moving closer to a “post-epistemological” world where “fact cannot be distinguished from fiction.”
Horowitz pointed to the prevalence of deep fakes and hallucinations. Hallucinations are phenomena This occurs when technology “recognizes patterns or objects that do not exist or are imperceptible to a human observer, creating output that is meaningless or totally inaccurate.”
It remains to be seen whether regulations and parameters for addressing AI deception will ever be revealed, but it is clear that if we want to combat AI-generated falsehoods, humans must establish a basis for truth.
Like Blaze News? Bypass the censorship and sign up for our newsletter to get stories like this directly to your inbox. Register here!
